Meta targets first-half 2027 deployment of its in-house MTIA 450 AI inference chip in data centers.

Meta Platforms plans to deploy its MTIA 450 custom AI chip across its data centers in the first half of 2027, according to Bloomberg. The processor is designed to reduce the cost and energy required to run AI models. The chip is Meta's third-generation offering, also known as Arke, and is currently undergoing testing. Design work on the next generation, MTIA 500 (codenamed Astrid), is expected to conclude within about a month, with broader deployment targeted for the end of 2027.

MTIA 450 has been engineered primarily for generative AI inference workloads, including the production of images and video from text prompts. The processor can also support recommendation systems and limited training tasks, though it is not intended for training large language models. According to Meta's March technical documentation, MTIA 450 delivers six times as many MX4 FLOPS compared to FP16/BF16 operations. Its mixed low-precision architecture is designed to maximize throughput while minimizing software conversion overhead and maintaining model fidelity.

MTIA 500 will offer approximately 50% greater high-bandwidth memory throughput than MTIA 450, along with 80% increased HBM capacity and 43% additional MX4 FLOPS. The design integrates four smaller compute chiplets with HBM stacks, network chiplets, and a system-on-chip component.

Meta has achieved a six-month development cycle between chip generations. The MTIA 400, 450, and 500 architectures share compatible chassis, rack, and network infrastructure, enabling Meta to deploy successive processors within the same physical footprint. "It's unusual for any silicon company or team to be releasing a new chip every six months. It's a very quick cadence," said Yee Jiun Song, Meta's Vice President of Engineering, in March. Song attributed this pace to Meta's rapid capacity expansion and capital spending, which make current-generation silicon valuable at each deployment point.

Meta is developing MTIA with Broadcom, while Taiwan Semiconductor Manufacturing Company produces the chips. Custom silicon provides Meta with greater control over price-to-performance and reduces reliance on a limited external supplier base. "This also provides us with more diversity in terms of silicon supply, and insulates us from price changes to some extent," Song said.

MTIA processors will operate alongside external accelerators. Meta has secured agreements covering millions of Nvidia GPUs and up to 6 gigawatts of AMD GPU deployments across multiple years. Song emphasized that rapidly evolving workloads necessitate maintaining multiple hardware options. The proliferation of later-generation MTIA chips will increase Meta's exposure to constrained HBM supply. Song acknowledged being "absolutely worried about HBM supply," though he expressed confidence that Meta had secured sufficient memory for its planned expansion.

Meta's chip roadmap accompanies expansion of its Hyperion campus in Louisiana to 5 gigawatts and a $14 billion partnership with BlackRock for a 1-gigawatt campus in El Paso, Texas. Meta has already deployed hundreds of thousands of MTIA processors across its production systems.